Design models
Models trained on public structures (the Protein Data Bank) and on our own lab data. They draw protein shapes for a target site and write sequences that should fold into them.
Platform
Our model designs new proteins. Our lab makes and tests them. Every result teaches the model, so each round is better than the last.
In many labs, a round from idea to lab result takes weeks. Our loop takes seven days.

The model reads the 3D shape of a site on the target. It draws new protein shapes that fit the site, then writes the sequences that should fold into them. It scores 100 million designs per target and picks a few thousand to make.
100Mdesigns scored per target

Robots build the DNA for each design and make the proteins in small batches with cell-free expression. A plate of 384 designs is ready in two days.
48 hfrom DNA to protein (cell-free)

Fast screens indicate which designs fold as drawn (CD and SEC), which bind the target (BLI) and which also bind similar proteins.
12,000designs screened each week

Every result, good or bad, goes back into the model. Each weekly round makes the next designs better.
4% → 27%designs with KD under 10 nM, round 1 → round 8
Why it matters
Small-molecule drugs usually need a pocket to hold on to, and many targets have none. Antibodies bind flat surfaces well, but they are large (about 150 kDa). Designed miniproteins are small (about 7 kDa), can be built for one exact site and can be made stable where the disease is.
Three parts that work as one system.
Models trained on public structures (the Protein Data Bank) and on our own lab data. They draw protein shapes for a target site and write sequences that should fold into them.
Two robot lines make and test proteins day and night. Every plate is tracked from DNA to result.
Every design, every test and every result lives in one place, ready for the next round of training.
We partner with drug companies on targets that are hard for their current tools. You bring the target; we design the key.